VideoGPA: Distilling Geometry Priors For 3D-Consistent Video Generation
- Posted by Jesse Polhemus
- on April 2, 2026
by Hongyang Du, Brown CS Master's student
I'm excited to share recent work from Professor Randall Balestriero's lab, developed in collaboration with the USC PSI Lab: VideoGPA: Distilling Geometry Priors for 3D-Consistent Video Generation.
In this paper, we leverage a 3D Geometric Foundation Model to build a self-supervised pipeline that evaluates 3D consistency in AI-generated videos. By integrating our video generation model with reinforcement learning, we are able to generate highly 3D-coherent and realistic videos. This approach significantly reduces morphing, flickering, and artifacts, outperforming current state-of-the-art methods.
Additionally, we designed an efficient, minimal data augmentation pipeline. When evaluated on complex real-world scenarios, we discovered that the model successfully generalizes from simple training environments to achieve excellent performance on hard cases. Excitingly, we also observed an emergence of motion coherence in dynamic scenes.
Our paper: https://arxiv.org/abs/2601.23286
Our website: https://hongyang-du.github.io/VideoGPA-Website/
Our code base: https://github.com/Hongyang-Du/VideoGPA
